Research On Domestic Garbage Detection Algorithm Based On Deep Learning

In recent years,with the rapid development of deep learning technology,various algorithms for object detection have emerged,which have been widely applied in military,transportation,security,and other fields.The core task of object detection is to accurately locate and classify the target.Currently,object detection algorithms used in garbage detection can achieve an accuracy rate of over 99%.However,in some complex scenes or small object detection scenarios,the accuracy of these algorithms is not satisfactory.Therefore,researching deep learning-based multi-object detection and classification algorithms,especially for complex scenes and small object detection,has significant practical significance.This thesis evaluates the performance of object detection algorithms using universal evaluation metrics and analyzes the detection effect of garbage detection algorithms.A garbage detection system is designed and developed based on the research work described in this thesis.(1)To address the problem of a small number and type of garbage detection datasets,mainly consisting of classification datasets,a series of methods were used to improve the quality and increasing the quantity of the dataset.Initially,a preliminary dataset was constructed by manual labeling based on the Huawei Cloud garbage dataset and partially collected data.Secondly,an improved Cycle GAN was used to transform the initial dataset images into garbage images with a nighttime style for data augmentation.Finally,the ROF image denoising technique was introduced to improve the quality and clarity of generated images.A garbage detection dataset containing 4major categories,44 minor categories,31,984 images,and 44,203 targets was constructed,providing a good data foundation for subsequent research.The m AP of Cycle GAN before and after improvement and data augmentation was compared,with an increase of 5.3% and 8.8%,respectively.(2)To address the issue of poor detection performance in nighttime style scenes and small object detection,YOLOv5 was improved,and the DST-YOLOv5 object detection algorithm was proposed.Firstly,the backbone benchmark network was improved through dual-channel feature extraction and weighted feature fusion to enhance the detection effect of the nighttime style dataset.Secondly,the Swin Transformer module and multi-scale detection head were introduced to enhance the global correlation of features and obtain shallow features to improve small object detection.Finally,experiments were conducted to compare DST-YOLOv5 with YOLOv5 before and after improvement and other currently popular object detection algorithms.The experimental results show that DST-YOLOv5 achieved 92.9 m AP and 172.3 FPS.(3)A garbage detection system was designed and developed,consisting of an image upload module,image processing module,data management module,and user interaction module.Firstly,functional and non-functional requirements were analyzed based on the actual application demands of garbage detection and classification tasks.Secondly,the overall architecture of the system and the design of system functions were completed according to the relevant requirements analysis,and DST-YOLOv5 was embedded as the garbage detection algorithm.Finally,the user interaction interface was implemented using Py Qt5 to provide engineering support for the implementation of the algorithm.